The results of polls done by different survey organizations often are slightly different and sometimes flatly disagree. But as critics of polling noted last week, a Harris Interactive poll released Wednesday was unique: a poll that disagreed with itself.

"Bush Leads by Eight Points -- or Two -- Depending on Definition of Likely Voters" proclaimed the headline on the Harris release announcing the results.

The headline -- and survey findings -- highlighted the frustrations facing pollsters this year as they attempt to fill their samples with people who actually will vote Nov. 2.

That is more difficult than it seems, and there is not one right answer, pollsters say. Most organizations use slightly different techniques to define who is likely to vote. The way individual polling firms define probable voters explains in large part the differences in the polls.

Moreover, these methods were first developed or refined in the low-turnout elections of the 1980s and 1990s. Many pollsters fear they may not be entirely accurate if turnout this year is exceptionally high, particularly among groups who in the past have been disproportionately absent from the polling booth, such as young adults and Latinos.

Even in the best of times, defining probable voters is a challenge. Ask people whether they are registered to vote, and 86 percent say they are, according to The Washington Post tracking poll. Further ask them, as The Post and most other polling organizations do, how likely they are to vote, and more than 80 percent say they're "absolutely certain." If true, that would mean turnout would top 70 percent on Nov. 2.

That is probably not going to happen. Turnout in 2000 was slightly greater than 51 percent. Curtis Gans, executive director of the Committee for the Study of the American Electorate, suspects it might be seven to 10 percentage points higher this year -- but not 20 points higher.

These inflated estimates of voting intention can skew survey results, because those who say they will vote may hold different candidate preferences than those who actually do vote. Studies have shown that the phantom voters tend to be Democrats.

To attempt to separate those who truly will vote in this presidential election from those who say they will but will not, pollsters rely on a variety of techniques.

The Post uses seven variables to define likely voters: whether the respondents state they are registered to vote, their intention to vote, voting history, interest in the presidential campaign, age, whether the respondents are voting for the first time in 2004 and whether they know the location of their polling place. These questions screen out about 40 percent of all adults ages 18 and older.

The Gallup Organization relies on a slightly different set of questions to identify probable voters, including a measure of enthusiasm, and then makes statistical adjustments to come up with its sample of likely voters.

The Gallup model has proved to be a reliable predictor of the final presidential vote on the eve of the election but has also tended to produce wild swings in candidate preference when applied in the weeks before a campaign.

These swings are produced not by real changes in preference but by the short-term rise and fall in levels of enthusiasm supporters feel about their respective candidates, according to political scientist Robert Erikson of Columbia University and his research colleagues, who analyzed the performance of the Gallup tracking poll in the 2000 presidential election.